The DARTH Workgroup
16 October 2019
Set of statistical methods that can handle:
Used for both inference and extrapolation
Survival: the probability of not experiencing an event unit some time \( t \)
\( S(t) = Pr(T >t) = 1- Pr(T \leq t) \)
Likelihood based metrics
Performance based metrics
Latimer et al 2011 (DSU 14)
Multi - parameter distributions
Generalized Gamma - Generalized F
Data hungry - possibiliy of overfitting
Hypothesis testing possible
Survival models where the underlying hazard of an event is modelled as a smooth, piecewise polynomial function of time.
Gibson et al (Pharmacoeconomics, 2017)
From Gibson et al (Pharacoeconomcs, 2017)
Spline-based models using a limited number of knots can provide an acceptable fit to trial data and generate extrapolated estimates supported by longer term evidence, with results that are stable in response to changes in knot placement.
Modeler faced with a more decisions:
Extrapolate using the fitted curve
Fit seperate curves to the data
Extrapolate treatment effect as relative
Extrapolate using KM and “fitted” tails
Drummond, et al (2015). Methods for the economic evaluation of health care programmes. Oxford university press.
Mechanics:
Implicit assumption: risk of dying is only a function of time
Less problematic with less censoring
Woods et al 2017 (DSU 19)
Williams et al 2017 (MDM)
Can incorporate:
Can work with
Fitted in two ways:
flexsurv and mstateFitted in two ways:
Drawbacks:
Can be fitted through msm and flexsurv
Great resource for both survival fitting and multistate models
Given short term follow up in RCTs, survival extrapolations can be informed by external data
Other sources of external data:
Observational data
{ width=20% }
Cure rate models a promising solution but the cure faction difficult to estimate
Can be fitted in both R and SAS
Guidance is lacking but given the popularity soon to come!